mdevora commited on
Commit
6f02abb
·
1 Parent(s): 31a0513

Create app.py

Browse files
Files changed (1) hide show
  1. app.py +174 -0
app.py ADDED
@@ -0,0 +1,174 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import itertools
2
+ import os
3
+ import gradio as gr
4
+ import openai
5
+ from langchain.chains import RetrievalQA
6
+ from langchain.document_loaders.directory import DirectoryLoader
7
+ from langchain.memory import ConversationBufferWindowMemory
8
+ from langchain.text_splitter import CharacterTextSplitter
9
+ import textwrap
10
+ from langchain.embeddings import OpenAIEmbeddings
11
+ from langchain.vectorstores import Chroma
12
+ from langchain_community.document_loaders import UnstructuredURLLoader
13
+ from langchain_core.callbacks import StreamingStdOutCallbackHandler
14
+ from langchain.prompts import PromptTemplate
15
+ from langchain.chat_models import ChatOpenAI
16
+ from pdfminer.high_level import extract_text
17
+
18
+
19
+ os.environ["HUGGINGFACEHUB_API_TOKEN"] = os.getenv('HUGGINGFACEHUB_API_TOKEN')
20
+ openai.api_key = os.getenv('OPENAI_API_KEY')
21
+ #loader = DirectoryLoader('documents/')
22
+ docs_dir = 'documents/'
23
+ txts_dir = docs_dir + 'txt/'
24
+ converted_txts_dir = docs_dir + 'converted_txt/'
25
+ # pdfs_dir = docs_dir + 'pdf/'
26
+ # txt_documents = [os.path.splitext(doc)[0] for doc in list(os.listdir(converted_txts_dir))]
27
+ # pdf_documents = [os.path.splitext(doc)[0] for doc in list(os.listdir(pdfs_dir))]
28
+ # for pdf_doc in pdf_documents:
29
+ # if pdf_doc not in txt_documents:
30
+ # print("Converting " + pdf_doc + ' pdf to txt')
31
+ # text = extract_text(pdfs_dir + pdf_doc + '.pdf')
32
+ # with open(converted_txts_dir + pdf_doc + '.txt', 'w') as f:
33
+ # f.write(text)
34
+ # intents_df = pd.read_csv(docs_dir + 'intents.csv')
35
+ # intents_df.sort_values(by=['intent', 'response'], inplace=True)
36
+ # previous_row = None
37
+ # for index, row in intents_df.iterrows():
38
+ # if previous_row is not None and previous_row['intent'] == row['intent']:
39
+ # intent_index = intent_index + 1
40
+ # else:
41
+ # intent_index = 0
42
+ # with open(f'{docs_dir}intents/{row["intent"]}_{intent_index}.txt', 'w') as f:
43
+ # f.write(row['response'])
44
+ # previous_row = row
45
+ docs_dir = 'documents/'
46
+ txts_dir = docs_dir + 'txt/'
47
+ # converted_txts_dir = docs_dir + 'converted_txt/'
48
+ # pdfs_dir = docs_dir + 'pdf/'
49
+ # txt_documents = [os.path.splitext(doc)[0] for doc in list(os.listdir(converted_txts_dir))]
50
+ # pdf_documents = [os.path.splitext(doc)[0] for doc in list(os.listdir(pdfs_dir))]
51
+ # for pdf_doc in pdf_documents:
52
+ # if pdf_doc not in txt_documents:
53
+ # print("Converting " + pdf_doc + ' pdf to txt')
54
+ # text = extract_text(pdfs_dir + pdf_doc + '.pdf')
55
+ # with open(converted_txts_dir + pdf_doc + '.txt', 'w') as f:
56
+ # f.write(text)
57
+ # intents_df = pd.read_csv(docs_dir + 'intents.csv')
58
+ # intents_df.sort_values(by=['intent', 'response'], inplace=True)
59
+ # previous_row = None
60
+ # for index, row in intents_df.iterrows():
61
+ # if previous_row is not None and previous_row['intent'] == row['intent']:
62
+ # intent_index = intent_index + 1
63
+ # else:
64
+ # intent_index = 0
65
+ # with open(f'{docs_dir}intents/{row["intent"]}_{intent_index}.txt', 'w') as f:
66
+ # f.write(row['response'])
67
+ # previous_row = row
68
+ loader = DirectoryLoader(txts_dir)
69
+ txt_documents = loader.load()
70
+ loader = UnstructuredURLLoader(urls=['https://stewartdavidpi.com/category/una-sociedad-magica/',
71
+ 'https://stewartdavidpi.com/category/hijos-de-kirves/',
72
+ 'https://stewartdavidpi.com/category/el-poder-tormerio/',
73
+ 'https://stewartdavidpi.com/category/el-senor-de-gon-dhul/',
74
+ 'https://stewartdavidpi.com/category/la-evolucion-tadur/',
75
+ 'https://stewartdavidpi.com/sobre-mi/',
76
+ 'https://stewartdavidpi.com/2023/12/01/sobre-las-ilustraciones-del-blog/',
77
+ 'https://stewartdavidpi.com/2023/11/17/mapa-de-inggrim/',
78
+ 'https://stewartdavidpi.com/2023/11/10/desarrollo-creativo-del-mapa-de-inggrim-parte-1-2/',
79
+ 'https://stewartdavidpi.com/2023/11/17/desarrollo-creativo-del-mapa-de-inggrim-2-2/',
80
+ 'https://stewartdavidpi.com/2023/11/03/tenemos-banda-sonora/'])
81
+ url_documents = loader.load()
82
+ def wrap_text_preserve_newlines(text, width=110):
83
+ # Split the input text into lines based on newline characters
84
+ lines = text.split('\n')
85
+ # Wrap each line individually
86
+ wrapped_lines = [textwrap.fill(line, width=width) for line in lines]
87
+ # Join the wrapped lines back together using newline characters
88
+ wrapped_text = '\n'.join(wrapped_lines)
89
+ return wrapped_text
90
+
91
+ max_length = 1000
92
+ text_splitter = CharacterTextSplitter(chunk_size=max_length, chunk_overlap=10)
93
+ docs = text_splitter.split_documents(itertools.chain(txt_documents, url_documents))
94
+
95
+ model_name = "decapoda-research/llama-7b-hf"
96
+ model_name = "gpt-3.5-turbo"
97
+ #model_name = "gpt2"
98
+ #tokenizer = GPT2Tokenizer.from_pretrained(model_name)
99
+ #tokenizer = tiktoken.get_encoding("gpt2")
100
+ #assert tokenizer.decode(tokenizer.encode("hola mundo")) == "hola mundo"
101
+ # tokenizer._pad_token = tokenizer.eos_token
102
+ # tokenizer._pad_token_id = 50256
103
+ #model = GPT2LMHeadModel.from_pretrained(model_name)
104
+ embeddings = OpenAIEmbeddings()
105
+ vector_store = Chroma.from_documents(docs, embeddings)
106
+ model_kwargs = {"max_length": max_length, "pad_token_id": 50256, "pad_token": "[PAD]", "padding": 'longest'}
107
+ # llm = HuggingFaceHub(repo_id=model_name, model_kwargs={"temperature": 0.1, "max_length": 500,
108
+ # "pad_token_id": 50256, "pad_token": "[PAD]",
109
+ # "padding": 'longest' #for open end generation
110
+ # })
111
+ llm = ChatOpenAI(streaming=True, model_name=model_name, temperature=0.9)
112
+ prompt_template = """You are going to
113
+ receive a context with information about a spanish writer and his sets of novels.
114
+ You can only answer things based on the question and a context to answer it.
115
+ Context: {context}
116
+ Question: {question}"""
117
+ #Memory: {chat_history}"""
118
+
119
+ PROMPT = PromptTemplate(
120
+ template=prompt_template, input_variables=["context", "question"]
121
+ )
122
+ chain_type_kwargs = {"prompt": PROMPT}
123
+ memory = ConversationBufferWindowMemory(k=4, memory_key="chat_history", input_key="question", output_key='answer',
124
+ return_messages=True)
125
+
126
+ retriever = vector_store.as_retriever()
127
+ # memory = VectorStoreRetrieverMemory(
128
+ # memory_key="chat_history",
129
+ # return_docs=False,
130
+ # retriever=retriever,
131
+ # return_messages=True,
132
+ # )
133
+ qa = RetrievalQA.from_chain_type(llm=llm, chain_type="stuff", retriever=retriever,
134
+ chain_type_kwargs=chain_type_kwargs, return_source_documents=True)
135
+ #memory=memory)
136
+ # qa = ConversationalRetrievalChain.from_llm(
137
+ # llm=llm, retriever=retriever, memory=memory,
138
+ # return_source_documents=True,
139
+ # combine_docs_chain_kwargs=chain_type_kwargs)
140
+
141
+
142
+ # while True:
143
+ # query = input("Question: ")
144
+ # #docs = vector_store.similarity_search(query)
145
+ # #print(wrap_text_preserve_newlines(str(docs[0].page_content)))
146
+ # response = qa({"query": query})
147
+ # print(f"Answer: {response['result']}")
148
+ # print(f"Sources: {response['source_documents']}")
149
+
150
+ def chat(message, history):
151
+ result = qa({"query": message})
152
+ sources_list = []
153
+ for item in result['source_documents']:
154
+ if 'http' in item.metadata['source']:
155
+ sources_list = sources_list + [item.metadata['source']]
156
+ sources_list = list(set(sources_list))
157
+ response = result['result']
158
+ if len(sources_list) > 0:
159
+ response = response + '\n\nSources:\n' + '\n'.join(sources_list)
160
+ yield response
161
+ # print(f"Answer: {response['result']}")
162
+ # print(f"Sources: {response['source_documents']}")
163
+
164
+ gr.ChatInterface(chat, chatbot=gr.Chatbot(),
165
+ textbox=gr.Textbox(placeholder="Pregúntame lo que quieras sobre INGGRIM!", container=False, scale=7),
166
+ title="Todo sobre la novela río: INGGRIM",
167
+ description="Pregúntame lo que quieras sobre INGGRIM!",
168
+ theme="soft",
169
+ examples=["¿Quién es Stewart David?", "¿Cuáles son las facciones presentes en INGGRIM?", "¿Cómo es el mapa del mundo de INGGRIM?"],
170
+ cache_examples=False,
171
+ retry_btn=None,
172
+ undo_btn="Borrar anterior",
173
+ clear_btn="Borrar").queue().launch()
174
+